A new approach to the development of artificial intelligence (AI) was proposed by scientists of Tyumen State University. According to specialists, with the help of the new model, AI will be “insured” from errors and capable of abstract thinking.
Robots and AI systems used nowadays derive their answers statistically based on an array of analyzed examples and can be effective in certain computational tasks, though so far they are far from “strong” AI systems. Therefore, according to scientists, it is not uncommon for those robots to make unethical decisions. As long as AI focuses on a given set of samples, without being able to expand it in almost real-time, there will be a risk that the automatic solution will be incorrect.
In order to reduce such risks, Russian specialists have proposed an approach to the development of AI, which will make it more intelligent and “conscious”. According to the authors of the study, their model of AI architecture combines the results of a wide range of physical, mathematical, cognitive, philosophical and other research in this field.
“We have deduced a key feature of the presence of a “common AI” system: it should be able to understand and be able to apply different models of its environment, that is, in other words, different ‘worldviews’ or ‘theories’. Focusing on ‘theories’ as certain paradigms of data evaluation, AI will not only be able to ‘look at the situation differently’ but also to deduce new principles, laws relevant to this area,” a researcher at the School of Advanced Studies of the university, Louis Vervoort, said.
Scientists explained that to think abstractly means to use consistent models and conceptual frameworks that allow studying the world from different angles, expanding interaction with it.
“The ethical problem with AI is not that machines do not have a common ‘human’ understanding of the world, which would include unambiguous moral positions and the idea of moral responsibility. High-level cognition, consciousness and ethical behaviour stem from one source – from the ability to master a broad conceptual framework and not just special tasks. We believe that this process can be reproduced in a machine,” Louis Vervoort said.
In the future, the research team plans to continue working in close cooperation with experts in the field of artificial intelligence and machine learning to develop specific computational approaches based on the new model. The article with the research results can be found in the Mind and Matter journal.
“So far we have proposed only a general strategy. But our analysis allows us to understand how much modern AI has advanced in the imitation or implementation of abstract thinking: if you look at the details, it turns out that ‘theoretical mastery’ is still far away. However, we believe that the future AI will be ‘controlled’ by theory, and not by data, as it is now,” Vervoort concluded.
Source: RIA Novosti